memory_consolidate (semantic) and memory_reflect timeout due to sequential KV operations

[ftr5672227]

Bug Description

memory_consolidate with tier: "semantic" and memory_reflect both time out consistently. The episodic tier works fine, and basic operations (memory_save,
memory_recall, memory_smart_search) all work.

Environment

• agentmemory version: 0.9.21
• Platform: Windows 11, Claude Code CLI
• LLM provider: Anthropic SDK → DeepSeek API (ANTHROPIC_BASE_URL=https://api.deepseek.com/anthropic)
• Claude Code default MCP tool timeout: 120 seconds

Root Cause Analysis

I traced through the bundled source (dist/index.mjs) and found 4 specific issues:

1. Sequential KV writes in semantic consolidate (lines ~7498-7540)

Each extracted <fact> from the LLM response triggers an individual await kv.set(), executed one at a time in a while loop. With 20+ facts, this accumulates significant
latency:

while ((match = factRegex.exec(response)) !== null) {
    // ...
    await kv.set(KV.semantic, existing.id, existing);  // sequential!
}

2. N+1 KV operations in reflect (lines ~12010-12102)

Reflect has nested loops: outer loop over concept clusters (up to 10-20), each making an LLM call, then inner loop over parsed insights doing await kv.get() + await kv.set() per
 insight — all sequential:

for (const conceptNames of conceptClusters) {        // 10-20 iterations
    const response = await provider.summarize(...);    // LLM call per cluster
    while ((match = insightRegex.exec(response)) !== null) {
        const existing = await kv.get(KV.insights, fp);   // sequential read
        await kv.set(KV.insights, insight.id, insight);    // sequential write
    }
}

10 clusters × 5 insights × 2 KV ops = 100 sequential awaits, easily exceeding timeout.

3. No explicit timeout on Anthropic SDK initialization (lines ~285-288)

this.client = new Anthropic({
    apiKey,
    ...baseURL ? { baseURL } : {}
    // no timeout specified — defaults to SDK's 60s
});

When using alternative base URLs (e.g., DeepSeek), complex summarization prompts can exceed the default timeout, causing a hard failure with no retry.

4. Inconsistent use of Promise.all()

The decay loop at line ~7620 already uses Promise.all() for batch KV writes, showing the pattern is known. But consolidate and reflect don't use it.

Suggested Fix

1. Batch KV writes — collect writes into an array, then await Promise.all(writes) (same pattern already used in the decay loop)
2. Add configurable timeout to the Anthropic client initialization, e.g., timeout: 120_000 or read from env
3. Parallelize cluster processing in reflect — process independent clusters with Promise.allSettled() instead of a sequential for-loop

Reproduction

# In Claude Code with agentmemory plugin installed
# 1. Accumulate some sessions worth of memory_save calls
# 2. Run:
memory_consolidate({ tier: "episodic" })   # ✅ works
memory_consolidate({ tier: "semantic" })   # ❌ timeout
memory_reflect({ maxClusters: 10 })        # ❌ timeout